PointNet in PyTorch

This is a PyTorch implementation of PointNet by Qi et al. published in CVPR 2017.

Summary of modules:

• PointNetFeatures (pointnet/model/pointnet.py): extracts point cloud features
• PointNetClassifier (pointnet/model/classifier.py): model for classification that builds upon PointNetFeatures
• PointSpatialTransformer (pointnet/model/stn.py): spatial transformer network for point clouds

Usage

python train_classifier.py ModelNet40 --dataset_dir /path/to/datasets --snapshot_dir /path/to/snapshots --snapshot_every 10 --with_validation --epochs 100
...

python test_classifier.py ModelNet40 /path/to/snapshots/best --dataset_dir /path/to/datasets

Testing PointNet Classifier
Dataset: /path/to/datasets/ModelNet40
Split: test
Snapshot: /path/to/snapshots/best

Accuracy: 84.64%, elapsed time: 15.410s, avg. time/batch: 0.002s